高美玲, 段锦, 赵伟强, 胡奇. 基于空洞全局注意力机制的近红外图像彩色化方法[J]. 红外技术, 2023, 45(10): 1096-1105.
引用本文: 高美玲, 段锦, 赵伟强, 胡奇. 基于空洞全局注意力机制的近红外图像彩色化方法[J]. 红外技术, 2023, 45(10): 1096-1105.
GAO Meiling, DUAN Jin, ZHAO Weiqiang, HU Qi. Near-infrared Image Colorization Method Based on a Dilated Global Attention Mechanism[J]. Infrared Technology , 2023, 45(10): 1096-1105.
Citation: GAO Meiling, DUAN Jin, ZHAO Weiqiang, HU Qi. Near-infrared Image Colorization Method Based on a Dilated Global Attention Mechanism[J]. Infrared Technology , 2023, 45(10): 1096-1105.

基于空洞全局注意力机制的近红外图像彩色化方法

Near-infrared Image Colorization Method Based on a Dilated Global Attention Mechanism

  • 摘要: 针对目前卷积神经网络未能充分提取图像的浅层特征信息导致近红外图像彩色化算法存在结果图像局部区域误着色及网络训练不稳定导致结果出现模糊问题,提出了一种新的生成对抗网络方法用于彩色化任务。首先,在生成器残差块中引入自行设计的空洞全局注意力模块,对近红外图像的每个位置理解更加充分,改善局部区域误着色问题;其次,在判别网络中,将批量归一化层替换成梯度归一化层,提升网络判别性能,改善彩色化图像生成过程带来的模糊问题;最后,将本文算法在RGB_NIR数据集上进行定性和定量对比。实验表明,本文算法与其他经典算法相比能充分提取近红外图像的浅层信息特征,在指标方面,结构相似性提高了0.044,峰值信噪比提高了0.835,感知相似度降低了0.021。

     

    Abstract: A new generative adversarial network method is proposed for colorization of near-infrared (NIR) images, because current convolutional neural networks fail to fully extract the shallow feature information of images. This failure leads to miscoloring of the local area of the resultant image and blurring due to unstable network training. First, a self-designed dilated global attention module was introduced into the generator residual block to identify each position of the NIR image accurately and improve the local region miscoloring problem. Second, in the discriminative network, the batch normalization layer was replaced with a gradient normalization layer to enhance the network discriminative performance and improve the blurring problem caused by the colorized image generation process. Finally, the algorithms used in this study are compared qualitatively and quantitatively using the RGB_NIR dataset. Experiments show that the proposed algorithm can fully extract the shallow information features of NIR images and improve the structural similarity by 0.044, PSNR by 0.835, and LPILS by 0.021 compared to other colorization algorithms.

     

/

返回文章
返回